AI research engineer working on LLM inference and serving: low-precision quantization (NVFP4), speculative decoding, and GPU serving on Kubernetes. Before that, 3 years in data science and data engineering.
- SGLang #35032: fix Llama 4 NVFP4 checkpoint loading (fused expert scales, packed q/k permute)
- vLLM #50773: showed the fusion passes match zero patterns on the NVFP4 path on sm_121 (v0.25, v0.26)
- agentic-rag-orchestrator: QLoRA-tuned Gemma-3-1B router that picks a RAG strategy per query. 98.7% routing accuracy, ~29% lower latency. M.Eng. thesis.
- DGX Spark serving stack (private): single-node kubeadm cluster on GB10, Helm charts for vLLM. Gemma-4-26B-A4B NVFP4 with MTP speculative decoding at 76-80 tok/s.
Blog · Seoul